Discriminative training of GMM based on Maximum Mutual Information for language identification

In this paper, a discriminative training procedures based on Maximum Mutual Information(MMI) for a Gaussian Mixture Model (GMM) language identification system is described. The idea is to find the model parameters lambda that minimize the conditional entropy H-lambda (C vertical bar X) of the random variable C given the random variable X, which means minimize the uncertainty in knowing what language was spoken given access to the utterance in X. The implementation of the proposal is based on the Generalized Probabilistic Descent (GPD) algorithm formulated to estimate the GMM parameters. The evaluation is conducted using the OGI multi-language telephone speech corpus. The experimental results show such system is very effective in language identification tasks.